A corporate AI hub running on the company’s own data
One way in to everything the company knows: operations, calls, advertising, tickets. An employee asks in plain language and gets the numbers, a summary and a conclusion they can act on. No queue to the dev team, and no data leaving the company.
01 · The problem
Decisions were made on gut feel
Prices, campaigns and priorities were set from managers’ experience and instinct rather than from numbers.
BI dashboards covered part of the picture, but every new question went straight back into the queue for the dev team.
Data lived in separate systems — operations, calls, advertising, tickets — and nobody had time to join it by hand.
Not everyone reads charts. A tool that demands an analyst’s skills only ever gets used by the analyst.
02 · What we did
We pulled the data together and opened it up through a chat
01Built the DWHInventory, requests and payments — the whole operation in a single ClickHouse warehouse.
02Transcribed five years of callsLocally, with Whisper: no paid telephony service and no recordings sent outside the company.
03Built a layer the model can readA DDS layer that documents and cleans every field, plus a context mart covering the business processes — so the model knows instead of guessing.
04Opened access through an MCP connectorTools across 6 data domains and 7 external APIs, a role-scoped link for every employee, and usage logs.
05Stood up the platform and rolled it outSelf-hosted LibreChat on the company’s own server, plus workshops with each department until people used it unaided.
03 · The result
Data became a working tool for every department
Marketing cut the campaigns that were not paying off by tying customer calls to Yandex Direct and site analytics.
Pricing and competitor analysis for a given item now take minutes and rest on demand data, not on instinct.
The product team brings hypotheses that already carry the supporting numbers and a prototype, so engineering receives work that is close to ready.
The IT team added tools of its own: logs and errors flow into the same warehouse and get triaged in the same chat.
04 · Technology
An open stack on our own hardware
Data never leaves the company perimeter: the server is owned outright and the cloud holds backups only. Against cloud VMs it pays for itself in 8–10 months.
Storage
ClickHouse · DWH · DDS layer
Connector
MCP on Laravel · domain tools · context mart
Speech and text
Whisper — call transcription, run locally
Interface
LibreChat (self-hosted), any LLM
Infrastructure
Own server, 2× AMD EPYC 7402 / 128 GB · Proxmox
External sources
Yandex Tracker · Metrica · AppMetrica · Direct · Webmaster · Wordstat
05 · What it looks like
Interface, architecture and infrastructure
05 · Stages
Stages of the work
Every stage closes with an artefact that stays with you — even if you carry on with your own team.
Data and risk audit2 weeks
DWH and the description layer3–5 weeks
MCP connector and access control3–4 weeks
Platform and rollout2–3 weeks
Sound familiar?
We start with a two-week audit: we look at what you can realistically hand to a model and say plainly where it will pay off in the first quarter and where it will not. The first conversation takes 30 minutes, no briefs or slide decks.